Research map / Image generation / Fast diffusion sampling
GLASS Flows: Efficient Inference for Reward Alignment of Flow and Diffusion Models
Peter Holderrieth, Uriel Singer, Tommi Jaakkola, Ricky T. Q. Chen, Yaron Lipman, Brian Karrer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace SDE transitions with GLASS Flows (rho = 0.4) in Feynman-Kac steering | FKS on FLUX · ImageReward, N=8 | ImageReward: 1.19 -> 1.40 | Improved |
| replace SDE transitions with GLASS Flows (rho = 0.4) in FKS | FKS on FLUX · ImageReward, N=8 | GenEval: 63.8 -> 74.3 | Improved |
| replace SDE transitions with GLASS Flows (rho = 0.4) in FKS | FKS on FLUX · CLIP reward, N=8 | CLIP reward: 39.0 -> 39.8 | Improved |
| replace SDE transitions with GLASS Flows (rho = 0.4) in FKS | FKS on FLUX · Pick reward, N=8 | GenEval: 64.1 -> 72.6 | Improved |
| use FKS with SDE instead of Best-of-N with SDE | SDE sampling · ImageReward, N=8 | GenEval: 65.3 -> 63.8 | No clear effect |
| use ODE instead of SDE for Best-of-N | Best-of-N · ImageReward, N=8 | ImageReward: 1.16 -> 1.31 | Improved |
| add SMC resampling (FKS) to GLASS Flows with rho = 0.4 (ablation) | GLASS Flows · ImageReward, N=8 | ImageReward: 1.32 -> 1.40 | Improved |
| set GLASS correlation rho = 0.4 instead of DDPM-equivalent rho (ablation) | FKS-GLASS · ImageReward, N=8 | ImageReward: 1.37 -> 1.40 | Improved |
| set GLASS correlation rho = 0.4 instead of DDPM-equivalent rho (ablation) | FKS-GLASS · ImageReward, N=8 | GenEval: 68.2 -> 74.3 | Improved |
| add reward gradient guidance to FKS-GLASS | FKS-GLASS · FLUX 672x672, ImageReward | ImageReward: 1.45 -> 1.52 | Improved |
| add reward gradient guidance to FKS-GLASS | FKS-GLASS · FLUX 672x672, ImageReward | GenEval: 72.7 -> 73.1 | Improved |
| use GLASS Flows instead of DDPM for posterior sampling at few simulation steps | posterior sampling · ImageNet256 SiT | FID | Improved |
Closest papers by meaning
- Discrete Flow Matching (NeurIPS 2024 · similarity 0.90)
- Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.90)
- PRECISE: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models (NeurIPS 2026 · similarity 0.90)
- Inductive Moment Matching (ICML 2025 · similarity 0.89)
- Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.89)
- Efficient Adjoint Matching for Fine-tuning Diffusion Models (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
